#include "mdadm.h"
#include "dlink.h"
#include <sys/mman.h>
+#include <stdint.h>
#if ! defined(__BIG_ENDIAN) && ! defined(__LITTLE_ENDIAN)
#error no endian defined
dprintf("Called restore_backup()\n");
fdlist = xmalloc(sizeof(int) * disk_count);
+ enable_fds(next_spare);
for (i = 0; i < next_spare; i++)
fdlist[i] = -1;
for (dev = content->devs; dev; dev = dev->next) {
if (fd < 0)
return;
- while (sysfs_fd_get_str(fd, action, 20) > 0 &&
- strncmp(action, "reshape", 7) == 0) {
- fd_set rfds;
- FD_ZERO(&rfds);
- FD_SET(fd, &rfds);
- select(fd+1, NULL, NULL, &rfds, NULL);
- }
+ while (sysfs_fd_get_str(fd, action, 20) > 0 &&
+ strncmp(action, "reshape", 7) == 0)
+ sysfs_wait(fd, NULL);
close(fd);
}
int d = 0;
struct mdinfo *sd;
+ enable_fds(nrdisks);
for (d = 0; d <= nrdisks; d++)
fdlist[d] = -1;
d = raid_disks;
return 1;
}
-unsigned long GCD(unsigned long a, unsigned long b)
-{
- while (a != b) {
- if (a < b)
- b -= a;
- if (b < a)
- a -= b;
- }
- return a;
-}
-
unsigned long compute_backup_blocks(int nchunk, int ochunk,
unsigned int ndata, unsigned int odata)
{
return blocks;
}
-char *analyse_change(struct mdinfo *info, struct reshape *re)
+char *analyse_change(char *devname, struct mdinfo *info, struct reshape *re)
{
/* Based on the current array state in info->array and
* the changes in info->new_* etc, determine:
/* chunk size is meaningful, must divide component_size
* evenly
*/
- if (info->component_size % (info->new_chunk/512))
- return "New chunk size does not"
- " divide component size";
+ if (info->component_size % (info->new_chunk/512)) {
+ unsigned long long shrink = info->component_size;
+ shrink &= ~(unsigned long long)(info->new_chunk/512-1);
+ pr_err("New chunk size (%dK) does not evenly divide device size (%lluk)\n",
+ info->new_chunk/1024, info->component_size/2);
+ pr_err("After shrinking any filesystem, \"mdadm --grow %s --size %llu\"\n",
+ devname, shrink/2);
+ pr_err("will shrink the array so the given chunk size would work.\n");
+ return "";
+ }
break;
default:
return "chunk size not meaningful for this level";
delta_parity = 1;
re->level = 5;
re->before.layout = ALGORITHM_PARITY_N;
+ if (info->new_layout == UnSet)
+ info->new_layout = map_name(r5layout, "default");
break;
case 6:
delta_parity = 2;
re->level = 6;
re->before.layout = ALGORITHM_PARITY_N;
+ if (info->new_layout == UnSet)
+ info->new_layout = map_name(r6layout, "default");
break;
default:
return "Impossible level change requested";
/* So we have a restripe operation, we need to calculate the number
* of blocks per reshape operation.
*/
+ re->new_size = info->component_size * re->before.data_disks;
if (info->new_chunk == 0)
info->new_chunk = info->array.chunk_size;
if (re->after.data_disks == re->before.data_disks &&
struct mdinfo info;
struct mdinfo *sra;
-
if (ioctl(fd, GET_ARRAY_INFO, &array) < 0) {
pr_err("%s is not an active md array - aborting\n",
devname);
if ((s->level == UnSet || s->level == array.level) &&
(s->layout_str == NULL) &&
(s->chunk == 0 || s->chunk == array.chunk_size) &&
+ data_offset == INVALID_SECTORS &&
(s->raiddisks == 0 || s->raiddisks == array.raid_disks)) {
/* Nothing more to do */
if (!changed && c->verbose >= 0)
return ret_val;
}
+static unsigned long long choose_offset(unsigned long long lo,
+ unsigned long long hi,
+ unsigned long long min,
+ unsigned long long max)
+{
+ /* Choose a new offset between hi and lo.
+ * It must be between min and max, but
+ * we would prefer something near the middle of hi/lo, and also
+ * prefer to be aligned to a big power of 2.
+ *
+ * So we start with the middle, then for each bit,
+ * starting at '1' and increasing, if it is set, we either
+ * add it or subtract it if possible, preferring the option
+ * which is furthest from the boundary.
+ *
+ * We stop once we get a 1MB alignment. As units are in sectors,
+ * 1MB = 2*1024 sectors.
+ */
+ unsigned long long choice = (lo + hi) / 2;
+ unsigned long long bit = 1;
+
+ for (bit = 1; bit < 2*1024; bit = bit << 1) {
+ unsigned long long bigger, smaller;
+ if (! (bit & choice))
+ continue;
+ bigger = choice + bit;
+ smaller = choice - bit;
+ if (bigger > max && smaller < min)
+ break;
+ if (bigger > max)
+ choice = smaller;
+ else if (smaller < min)
+ choice = bigger;
+ else if (hi - bigger > smaller - lo)
+ choice = bigger;
+ else
+ choice = smaller;
+ }
+ return choice;
+}
+
static int set_new_data_offset(struct mdinfo *sra, struct supertype *st,
char *devname, int delta_disks,
unsigned long long data_offset,
- unsigned long long min)
+ unsigned long long min,
+ int can_fallback)
{
struct mdinfo *sd;
int dir = 0;
int err = 0;
+ unsigned long long before, after;
+ /* Need to find min space before and after so same is used
+ * on all devices
+ */
+ before = UINT64_MAX;
+ after = UINT64_MAX;
for (sd = sra->devs; sd; sd = sd->next) {
char *dn;
int dfd;
/* Metadata doesn't support data_offset changes */
return 1;
}
+ if (before > info2.space_before)
+ before = info2.space_before;
+ if (after > info2.space_after)
+ after = info2.space_after;
+
+ if (data_offset != INVALID_SECTORS) {
+ if (dir == 0) {
+ if (info2.data_offset == data_offset) {
+ pr_err("%s: already has that data_offset\n",
+ dn);
+ goto release;
+ }
+ if (data_offset < info2.data_offset)
+ dir = -1;
+ else
+ dir = 1;
+ } else if ((data_offset <= info2.data_offset && dir == 1) ||
+ (data_offset >= info2.data_offset && dir == -1)) {
+ pr_err("%s: differing data offsets on devices make this --data-offset setting impossible\n",
+ dn);
+ goto release;
+ }
+ }
+ }
+ if (before == UINT64_MAX)
+ /* impossible really, there must be no devices */
+ return 1;
+
+ for (sd = sra->devs; sd; sd = sd->next) {
+ char *dn = map_dev(sd->disk.major, sd->disk.minor, 0);
+ unsigned long long new_data_offset;
+
+ if (sd->disk.state & (1<<MD_DISK_FAULTY))
+ continue;
if (delta_disks < 0) {
/* Don't need any space as array is shrinking
* just move data_offset up by min
*/
if (data_offset == INVALID_SECTORS)
- info2.new_data_offset = info2.data_offset + min;
+ new_data_offset = sd->data_offset + min;
else {
- if (data_offset < info2.data_offset + min) {
+ if (data_offset < sd->data_offset + min) {
pr_err("--data-offset too small for %s\n",
dn);
goto release;
}
- info2.new_data_offset = data_offset;
+ new_data_offset = data_offset;
}
} else if (delta_disks > 0) {
/* need space before */
- if (info2.space_before < min) {
+ if (before < min) {
+ if (can_fallback)
+ goto fallback;
pr_err("Insufficient head-space for reshape on %s\n",
dn);
goto release;
}
if (data_offset == INVALID_SECTORS)
- info2.new_data_offset = info2.data_offset - min;
+ new_data_offset = sd->data_offset - min;
else {
- if (data_offset > info2.data_offset - min) {
+ if (data_offset > sd->data_offset - min) {
pr_err("--data-offset too large for %s\n",
dn);
goto release;
}
- info2.new_data_offset = data_offset;
+ new_data_offset = data_offset;
}
} else {
if (dir == 0) {
- /* can move up or down. 'data_offset'
- * might guide us, otherwise choose
- * direction with most space
+ /* can move up or down. If 'data_offset'
+ * was set we would have already decided,
+ * so just choose direction with most space.
*/
- if (data_offset == INVALID_SECTORS) {
- if (info2.space_before > info2.space_after)
- dir = -1;
- else
- dir = 1;
- } else if (data_offset < info2.data_offset)
+ if (before > after)
dir = -1;
else
dir = 1;
- sysfs_set_str(sra, NULL, "reshape_direction",
- dir == 1 ? "backwards" : "forwards");
}
- switch (dir) {
- case 1: /* Increase data offset */
- if (info2.space_after < min) {
+ sysfs_set_str(sra, NULL, "reshape_direction",
+ dir == 1 ? "backwards" : "forwards");
+ if (dir > 0) {
+ /* Increase data offset */
+ if (after < min) {
+ if (can_fallback)
+ goto fallback;
pr_err("Insufficient tail-space for reshape on %s\n",
dn);
goto release;
}
if (data_offset != INVALID_SECTORS &&
- data_offset < info2.data_offset + min) {
+ data_offset < sd->data_offset + min) {
pr_err("--data-offset too small on %s\n",
dn);
goto release;
}
if (data_offset != INVALID_SECTORS)
- info2.new_data_offset = data_offset;
- else {
- unsigned long long off =
- info2.space_after / 2;
- off &= ~7ULL;
- if (off < min)
- off = min;
- info2.new_data_offset =
- info2.data_offset + off;
- }
- break;
- case -1: /* Decrease data offset */
- if (info2.space_before < min) {
+ new_data_offset = data_offset;
+ else
+ new_data_offset = choose_offset(sd->data_offset,
+ sd->data_offset + after,
+ sd->data_offset + min,
+ sd->data_offset + after);
+ } else {
+ /* Decrease data offset */
+ if (before < min) {
+ if (can_fallback)
+ goto fallback;
pr_err("insufficient head-room on %s\n",
dn);
goto release;
}
if (data_offset != INVALID_SECTORS &&
- data_offset < info2.data_offset - min) {
+ data_offset < sd->data_offset - min) {
pr_err("--data-offset too small on %s\n",
dn);
goto release;
}
if (data_offset != INVALID_SECTORS)
- info2.new_data_offset = data_offset;
- else {
- unsigned long long off =
- info2.space_before / 2;
- off &= ~7ULL;
- if (off < min)
- off = min;
- info2.new_data_offset =
- info2.data_offset - off;
- }
- break;
+ new_data_offset = data_offset;
+ else
+ new_data_offset = choose_offset(sd->data_offset - before,
+ sd->data_offset,
+ sd->data_offset - before,
+ sd->data_offset - min);
}
}
- if (sysfs_set_num(sra, sd, "new_offset",
- info2.new_data_offset) < 0) {
- err = errno;
- err = -1;
+ err = sysfs_set_num(sra, sd, "new_offset", new_data_offset);
+ if (err < 0 && errno == E2BIG) {
+ /* try again after increasing data size to max */
+ err = sysfs_set_num(sra, sd, "size", 0);
+ if (err < 0 && errno == EINVAL &&
+ !(sd->disk.state & (1<<MD_DISK_SYNC))) {
+ /* some kernels have a bug where you cannot
+ * use '0' on spare devices. */
+ sysfs_set_num(sra, sd, "size",
+ (sra->component_size + after)/2);
+ }
+ err = sysfs_set_num(sra, sd, "new_offset",
+ new_data_offset);
+ }
+ if (err < 0) {
+ if (errno == E2BIG && data_offset != INVALID_SECTORS) {
+ pr_err("data-offset is too big for %s\n",
+ dn);
+ goto release;
+ }
if (sd == sra->devs &&
(errno == ENOENT || errno == E2BIG))
/* Early kernel, no 'new_offset' file,
return err;
release:
return -1;
+fallback:
+ /* Just use a backup file */
+ return 1;
}
static int raid10_reshape(char *container, int fd, char *devname,
}
}
err = set_new_data_offset(sra, st, devname, info->delta_disks, data_offset,
- min);
+ min, 0);
if (err == 1) {
pr_err("Cannot set new_data_offset: RAID10 reshape not\n");
cont_err("supported on this kernel\n");
return -1;
}
+static int impose_level(int fd, int level, char *devname, int verbose)
+{
+ char *c;
+ struct mdu_array_info_s array;
+ struct mdinfo info;
+ sysfs_init(&info, fd, NULL);
+
+ ioctl(fd, GET_ARRAY_INFO, &array);
+ if (level == 0 &&
+ (array.level >= 4 && array.level <= 6)) {
+ /* To convert to RAID0 we need to fail and
+ * remove any non-data devices. */
+ int found = 0;
+ int d;
+ int data_disks = array.raid_disks - 1;
+ if (array.level == 6)
+ data_disks -= 1;
+ if (array.level == 5 &&
+ array.layout != ALGORITHM_PARITY_N)
+ return -1;
+ if (array.level == 6 &&
+ array.layout != ALGORITHM_PARITY_N_6)
+ return -1;
+ sysfs_set_str(&info, NULL,"sync_action", "idle");
+ /* First remove any spares so no recovery starts */
+ for (d = 0, found = 0;
+ d < MAX_DISKS && found < array.nr_disks;
+ d++) {
+ mdu_disk_info_t disk;
+ disk.number = d;
+ if (ioctl(fd, GET_DISK_INFO, &disk) < 0)
+ continue;
+ if (disk.major == 0 && disk.minor == 0)
+ continue;
+ found++;
+ if ((disk.state & (1 << MD_DISK_ACTIVE))
+ && disk.raid_disk < data_disks)
+ /* keep this */
+ continue;
+ ioctl(fd, HOT_REMOVE_DISK,
+ makedev(disk.major, disk.minor));
+ }
+ /* Now fail anything left */
+ ioctl(fd, GET_ARRAY_INFO, &array);
+ for (d = 0, found = 0;
+ d < MAX_DISKS && found < array.nr_disks;
+ d++) {
+ int cnt;
+ mdu_disk_info_t disk;
+ disk.number = d;
+ if (ioctl(fd, GET_DISK_INFO, &disk) < 0)
+ continue;
+ if (disk.major == 0 && disk.minor == 0)
+ continue;
+ found++;
+ if ((disk.state & (1 << MD_DISK_ACTIVE))
+ && disk.raid_disk < data_disks)
+ /* keep this */
+ continue;
+ ioctl(fd, SET_DISK_FAULTY,
+ makedev(disk.major, disk.minor));
+ cnt = 5;
+ while (ioctl(fd, HOT_REMOVE_DISK,
+ makedev(disk.major, disk.minor)) < 0
+ && errno == EBUSY
+ && cnt--) {
+ usleep(10000);
+ }
+ }
+ }
+ c = map_num(pers, level);
+ if (c) {
+ int err = sysfs_set_str(&info, NULL, "level", c);
+ if (err) {
+ err = errno;
+ pr_err("%s: could not set level to %s\n",
+ devname, c);
+ if (err == EBUSY &&
+ (array.state & (1<<MD_SB_BITMAP_PRESENT)))
+ cont_err("Bitmap must be removed"
+ " before level can be changed\n");
+ return err;
+ }
+ if (verbose >= 0)
+ pr_err("level of %s changed to %s\n",
+ devname, c);
+ }
+ return 0;
+}
+
static int reshape_array(char *container, int fd, char *devname,
struct supertype *st, struct mdinfo *info,
int force, struct mddev_dev *devlist,
info->new_level = UnSet;
if (info->delta_disks > 0)
info->array.raid_disks -= info->delta_disks;
- msg = analyse_change(info, &reshape);
+ msg = analyse_change(devname, info, &reshape);
info->new_level = new_level;
if (info->delta_disks > 0)
info->array.raid_disks += info->delta_disks;
/* Make sure the array isn't read-only */
ioctl(fd, RESTART_ARRAY_RW, 0);
} else
- msg = analyse_change(info, &reshape);
+ msg = analyse_change(devname, info, &reshape);
if (msg) {
- pr_err("%s\n", msg);
+ /* if msg == "", error has already been printed */
+ if (msg[0])
+ pr_err("%s\n", msg);
goto release;
}
if (restart &&
/* reshape already started. just skip to monitoring the reshape */
if (reshape.backup_blocks == 0)
return 0;
+ if (restart & RESHAPE_NO_BACKUP)
+ return 0;
+
+ /* Need 'sra' down at 'started:' */
+ sra = sysfs_read(fd, NULL,
+ GET_COMPONENT|GET_DEVS|GET_OFFSET|GET_STATE|GET_CHUNK|
+ GET_CACHE);
+ if (!sra) {
+ pr_err("%s: Cannot get array details from sysfs\n",
+ devname);
+ goto release;
+ }
goto started;
}
/* The container is frozen but the array may not be.
}
if (reshape.level != array.level) {
- char *c = map_num(pers, reshape.level);
- int err;
- if (c == NULL)
+ int err = impose_level(fd, reshape.level, devname, verbose);
+ if (err)
goto release;
-
- err = sysfs_set_str(info, NULL, "level", c);
- if (err) {
- err = errno;
- pr_err("%s: could not set level to %s\n",
- devname, c);
- if (err == EBUSY &&
- (info->array.state & (1<<MD_SB_BITMAP_PRESENT)))
- cont_err("Bitmap must be removed"
- " before level can be changed\n");
- goto release;
- }
- if (verbose >= 0)
- pr_err("level of %s changed to %s\n",
- devname, c);
+ info->new_layout = UnSet; /* after level change,
+ * layout is meaningless */
orig_level = array.level;
sysfs_freeze_array(info);
Manage_subdevs(devname, fd, devlist, verbose,
0,NULL, 0);
+ if (reshape.backup_blocks == 0 && data_offset != INVALID_SECTORS)
+ reshape.backup_blocks = reshape.before.data_disks * info->array.chunk_size/512;
if (reshape.backup_blocks == 0) {
/* No restriping needed, but we might need to impose
* some more changes: layout, raid_disks, chunk_size
goto release;
}
-started:
-
if (array.level == 10) {
/* Reshaping RAID10 does not require any data backup by
* user-space. Instead it requires that the data_offset
}
if (!backup_file)
- switch(set_new_data_offset(sra, st, devname, info->delta_disks,
+ switch(set_new_data_offset(sra, st, devname,
+ reshape.after.data_disks - reshape.before.data_disks,
data_offset,
- reshape.min_offset_change)) {
+ reshape.min_offset_change, 1)) {
case -1:
goto release;
case 0:
pr_err("Failed to initiate reshape!\n");
goto release;
}
-
+ if (info->new_level == reshape.level)
+ return 0;
+ /* need to adjust level when reshape completes */
+ switch(fork()) {
+ case -1: /* ignore error, but don't wait */
+ return 0;
+ default: /* parent */
+ return 0;
+ case 0:
+ map_fork();
+ break;
+ }
+ close(fd);
+ wait_reshape(sra);
+ fd = open_dev(sra->sys_name);
+ if (fd >= 0)
+ impose_level(fd, info->new_level, devname, verbose);
return 0;
case 1: /* Couldn't set data_offset, try the old way */
if (data_offset != INVALID_SECTORS) {
break;
}
+started:
/* Decide how many blocks (sectors) for a reshape
* unit. The number we have so far is just a minimum
*/
if (backup_file == NULL) {
if (reshape.after.data_disks <=
reshape.before.data_disks) {
- pr_err("%s: Cannot grow - "
- "need backup-file\n", devname);
+ pr_err("%s: Cannot grow - need backup-file\n",
+ devname);
+ pr_err(" Please provide one with \"--backup=...\"\n");
goto release;
} else if (sra->array.spare_disks == 0) {
pr_err("%s: Cannot grow - "
do {
struct mdstat_ent *mds, *m;
delayed = 0;
- mds = mdstat_read(0, 0);
+ mds = mdstat_read(1, 0);
for (m = mds; m; m = m->next)
if (strcmp(m->devnm, sra->sys_name) == 0) {
if (m->resync &&
delayed = 0;
}
if (delayed)
- sleep(30 - (delayed-1) * 25);
+ mdstat_wait(30 - (delayed-1) * 25);
} while (delayed);
-
+ mdstat_close();
close(fd);
if (check_env("MDADM_GROW_VERIFY"))
fd = open(devname, O_RDONLY | O_DIRECT);
set_array_size(st, info, info->text_version);
if (info->new_level != reshape.level) {
-
- c = map_num(pers, info->new_level);
- if (c) {
- err = sysfs_set_str(sra, NULL, "level", c);
- if (err)
- pr_err("%s: could not set level "
- "to %s\n", devname, c);
- }
+ if (fd < 0)
+ fd = open(devname, O_RDONLY);
+ impose_level(fd, info->new_level, devname, verbose);
+ close(fd);
if (info->new_level == 0)
st->update_tail = NULL;
}
flush_mdmon(container);
rv = reshape_array(container, fd, adev, st,
- content, force, NULL, 0ULL,
+ content, force, NULL, INVALID_SECTORS,
backup_file, verbose, 1, restart,
freeze_reshape);
close(fd);
unsigned long long backup_point,
unsigned long long wait_point,
unsigned long long *suspend_point,
- unsigned long long *reshape_completed)
+ unsigned long long *reshape_completed, int *frozen)
{
/* This function is called repeatedly by the reshape manager.
* It determines how much progress can safely be made and allows
wait_point = info->component_size - wait_point;
}
- sysfs_set_num(info, NULL, "sync_max", max_progress);
+ if (!*frozen)
+ sysfs_set_num(info, NULL, "sync_max", max_progress);
/* Now wait. If we have already reached the point that we were
* asked to wait to, don't wait at all, else wait for any change.
* waiting forever on a dead array
*/
char action[20];
- fd_set rfds;
if (sysfs_get_str(info, NULL, "sync_action",
action, 20) <= 0 ||
strncmp(action, "reshape", 7) != 0)
&& info->reshape_progress < (info->component_size
* reshape->after.data_disks))
break;
- FD_ZERO(&rfds);
- FD_SET(fd, &rfds);
- select(fd+1, NULL, NULL, &rfds, NULL);
+ sysfs_wait(fd, NULL);
if (sysfs_fd_get_ll(fd, &completed) < 0)
goto check_progress;
}
/* The abort might only be temporary. Wait up to 10
* seconds for fd to contain a valid number again.
*/
- struct timeval tv;
+ int wait = 10000;
int rv = -2;
- tv.tv_sec = 10;
- tv.tv_usec = 0;
- while (fd >= 0 && rv < 0 && tv.tv_sec > 0) {
- fd_set rfds;
- FD_ZERO(&rfds);
- FD_SET(fd, &rfds);
- if (select(fd+1, NULL, NULL, &rfds, &tv) != 1)
+ unsigned long long new_sync_max;
+ while (fd >= 0 && rv < 0 && wait > 0) {
+ if (sysfs_wait(fd, &wait) != 1)
break;
switch (sysfs_fd_get_ll(fd, &completed)) {
case 0:
/* all good again */
rv = 1;
+ /* If "sync_max" is no longer max_progress
+ * we need to freeze things
+ */
+ sysfs_get_ll(info, NULL, "sync_max", &new_sync_max);
+ *frozen = (new_sync_max != max_progress);
break;
case -2: /* read error - abort */
- tv.tv_sec = 0;
+ wait = 0;
break;
}
}
struct mdinfo *sd;
unsigned long stripes;
int uuid[4];
+ int frozen = 0;
/* set up the backup-super-block. This requires the
* uuid from the array.
wait_point = __le64_to_cpu(bsb.arraystart2);
}
+ reshape_completed = sra->reshape_progress;
rv = progress_reshape(sra, reshape,
backup_point, wait_point,
- &suspend_point, &reshape_completed);
+ &suspend_point, &reshape_completed,
+ &frozen);
/* external metadata would need to ping_monitor here */
sra->reshape_progress = reshape_completed;
close(cfd);
ret_val = reshape_container(st->container_devnm, NULL, mdfd,
st, info, 0, backup_file,
- 0, 1, freeze_reshape);
+ 0,
+ 1 | info->reshape_active,
+ freeze_reshape);
} else
ret_val = reshape_array(NULL, mdfd, "array", st, info, 1,
- NULL, 0ULL, backup_file, 0, 0, 1,
+ NULL, INVALID_SECTORS,
+ backup_file, 0, 0,
+ 1 | info->reshape_active,
freeze_reshape);
return ret_val;